Environmental Engineering Reference
In-Depth Information
Precise
Imprecise
True
Unbiased
Biased
Fig. 2.1
When estimating a parameter such as abundance, there are four possible
outcomes, defined by precision and bias that are either high or low. Parameter
estimates are indicated by points, with the degree of statistical confidence in each
estimate indicated by error bars. The true parameter value is indicated by the
dashed line.
we are looking for methods that minimise bias and maximise precision, but of
these two factors, low bias is the more important (Figure 2.1). This is because
actions based on a biased estimate can be disastrous (e.g. a decision to continue
harvest of a depleted population because of a mistakenly high population
estimate), whereas low precision should simply increase the caution with which
results are treated. We describe below the principles of bias and precision, and
outline how to achieve the best possible sample in practice.
Sampling requires the definition and selection of
sampling units
. For example,
these may be discrete sites for estimating abundance (Section 2.3), individual
organisms for estimating demographic rates (Section 2.4), or individuals, house-
holds or communities of people when studying resource users (Chapter 3). The
ways in which sampling units are defined and selected have important implications
for bias and precision, which we outline in the following sections.
A key source of bias is the failure to select
representative
sampling units. For
example, haphazard selection (throwing quadrats, sticking pins in maps, selecting
individuals on encounter, etc.) is not adequate because it allows subconscious
selection of sites that 'feel right', or can be influenced by variation in detectability
between individual organisms. Either case tends to result in an unrepresentative
sample.
A second important source of bias is the failure to meet the
assumptions
of the
analysis. Such inappropriate use of a model gives rise to
model error
, which can
result in serious bias. Most of the methods described in this chapter not only
sample many sites or individuals, but also use some form of model to estimate
parameters. It is crucial that the assumptions of these analyses are understood in
